EPSRC Centre for Doctoral Training in Industrially Focused ...
Transcript of EPSRC Centre for Doctoral Training in Industrially Focused ...
EPSRC Centre for Doctoral Training in
Industrially Focused Mathematical
Modelling
Distribution of signal-to-interference in
cellular wireless networks
Fabian Ying
1
Contents1. Introduction .............................................2
Background ..................................................2
2. Mathematical model...............................2
Glossary of terms........................................4
Comments ....................................................4
3. Comparison of PPP model with using
house locations for femtocell locations .....4
4. Comparison with commercial software .5
5. Inter-cell interference coordination ......6
6. Conclusion and recommendations ..........7
7. Potential Impact .....................................7
References ...................................................7
2
1. Introduction
BackgroundA new way of connecting mobile devices to cellular networks are femtocells (see Figure 1),
which are small, low-power cellular base stations, typically distributed in homes and
businesses. If deployed, network operators such as BT are likely to incorporate them in
WiFi access points. The femtocells will provide additional cellular coverage to users,
especially indoors, by filling gaps of coverage areas of the main network and eliminating
loss of signal inside buildings. Furthermore, the capacity of the main network is improved
as traffic is off-loaded through the user’s network (internet) via the femtocells, so fewer
users need to connect to the main network, which is done traditionally via cellular towers
(called macrocells). Network operators in the UK are considering the deployment of home
femtocells, and are therefore investigating the potential performance of such systems.
In this report, we consider the problem of finding the distribution of the signal-to-interference
ratio (SIR) in a femtocell network. The SIR of a particular user equipment (UE), such as a
mobile phone, is the ratio between the received signal power from the serving cellular base
station and the total interference power from all other base stations which transmit on the
same frequency band. SIR is a key performance indicator in cellular wireless networks, as it
determines, using Shannon’s law, the maximal achievable datarate that a user can receive.
Our aim is to find the distribution via the complementary cumulative distribution function
(ccdf) of SIR in wireless networks which gives us the proportion of users that can achieve
a given input SIR on a single frequency band.
Figure 1: Model of a BT Smart Hub. Femtocells, if deployed, will be similar devices.
In this project, we use simplified mathematical models from the literature for general
cellular wireless networks and apply them to the femtocell case. When tractable analytical
solutions to the simplified mathematical problem exist, they provide quick approximations
to realistic scenarios, making it possible to test out different scenarios without having to
rely on slow and expensive commercial software to perform complex simulations or
conduct field tests which are both expensive and prone to inaccuracy due to noise. In our
work, we test how well the mathematical model agrees with more realistic data as well as
data from commercial software, and apply the mathematical model to a system with inter-
cell interference coordination (ICIC).
2. Mathematical modelThe mathematical model for the cellular network that we use consists of a random model
for the spatial locations of the femtocells (cells) coupled with a standard power loss
propagation model with random channel effects.
Simplifiedmathematical modelsare useful to quicklytest out differenttransmissionscenarios withouthaving to rely onslow, complexsimulations.
SIR is a key quantity incellular networks, as itdetermines thedatarate that a usercan receive.
3
We assume that the locations of base stations are randomly distributed and we use a
homogeneous Poisson Point Process (PPP) to provide our spatial model (see Figure 2 for
one realisation of a PPP). Since the exact locations of femtocells will often be decided by
the user and not by the network provider, it is reasonable to assume the positions of the
femtocells are random. However, the PPP model has two features that likely do not occur
in practice: First, there is no minimum distance between points in the PPP model, resulting
in clusters of points which typically would not occur in a femtocell deployment. Second,
we assume that the density of points in the PPP model is constant in space, which is likely
not going to be true in reality due to parks or lakes which cannot contain any femtocells.
Despite these two shortcomings, the PPP model is widely used in the wireless
communication literature due to its mathematical tractability. Finally, we assume in our
model that UEs are distributed uniformly at random, and that each UE connects to the
nearest cell.
Figure 2: One realisation of a PPP (red dots) with Voronoi cell boundaries (blue curves).The Voronoi cell of a point is the region in which the given point is closer than any otherpoint. Hence all UEs in a Voronoi cell connect to the associated PPP point (red dot).
In a wireless network, attenuation of the radiated signals from cells to the UE can be
caused by propagation effects. The main three propagation effects are (i) pathloss (signals
lose power with distance), (ii) multipath fading (reception of multiple copies of the same
signal), and (iii) shadowing (blocking of the signal due to large obstacles such as buidling). In
our propagation model, we include includes pathloss and Rayleigh fading (a specific model
of multipath fading), but we omit shadowing for the sake of mathematical tractability.
Furthermore, we assume that all cells transmit at the same power.
This mathematical model has been widely studied in the literature. Andrews et al. [1]
derived a tractable analytical formula for the ccdf of SIR for the PPP model with pathloss
and Rayleigh fading (the baseline model described above). In this report, we will compare
this analytical formula with results from simple simulations of more realistic scenarios and
data from complex simulations using commercial software. It has been observed in the
literature that the ccdf of SIR in other spatial or propagation models is merely a
horizontally shifted version of the SIR curve from the PPP model. Ganti and Haenggi [2]
derived analytical expressions for estimating this shift (or gap) to the ccdf of the SIR from
the baseline model. They derived two approximations, one based on the gap on the left
asymptotic end (for low SIR values) and the other one based right asymptotic end (for high
SIR values). While the computation of the gap estimation approximations involves
Our mathematicalmodel consists of arandom model for celllocations and astandard propagationmodel for the signal.
Our baseline model isthe PPP model for celllocations with signalsexperiencing pathlossand Rayleigh fading.
4
simulations as well, it gives an approximation of the entire shape of the ccdf of SIR,
including the probability of rare events, whereas direct simulations exhibit much larger
errors for estimating the probability of rare events. We use both approximations in our
work and call the approximation ‘gap estimation’.
Glossary of terms
UE: User equipment (e.g. mobile phone).
Cell: Cellular base station (femtocells in our case).
SIR: Signal-to-interference ratio =received signal power
total interference power.
ccdf: Complementary cumulative distribution function. The ccdf of SIR gives theproportion of UEs that can achieve a given SIR, and hence a datarate.
PPP model: Model for locations of cell location, used because of its mathematicaltractability.
Baseline model: Our baseline model is a PPP model for cell location with pathlossand Rayleigh fading.
Comments• Both the spatial and the propagation model use many assumptions for the sake of
mathematical tractability and these assumptions may not be true in real-worldscenarios. However, mathematical models that capture the important qualitativefeatures can give good approximation of the real system.
• The propagation model is widely used in the literature and adapted versions of itare often used in commercial software.
3. Comparison of PPP model with using house
locations for femtocell locationsWe aim to validate the predictions of the ccdf of SIR of the PPP model with Rayleigh
fading (using the analytical results from [1]) by changing the spatial model and instead
using a more realistic dataset for femtocell locations. Note that we do not alter the
propagation model. As discussed above, PPP points may not be a realistic representation
of femtocell locations, due to the clustering of points occurring in a PPP. Hence we use
house locations data from Ordnance Survey OpenData (OS Open Map Local) and assume
that there is one femtocell located in each house (100% penetration case). We ran
simulations to compute the ccdf of SIR in urban, suburban, and rural areas using the
standard propagation model described in Section 2 with pathloss and Rayleigh fading and
with UEs distributed randomly. We compare the results with the analytical formula for the
ccdf of SIR of the PPP model and also apply the two approximation using gap estimation
(see Figure 3). In our simulations, we observed that the ccdf of SIR in urban areas agrees
very well with the PPP model (e.g. see Figure 3a). On the other hand, the ccdf of SIR (blue
dotted curve in Figure 3b) in suburban and rural areas show significant discrepancies to the
ccdf of SIR (red curve in Figure 3b) of PPP model, displaying worse performance than the
PPP model would have predicted. However, gap estimation (in particular, the
approximation using the left asymptotic gap (see Figure 3b)) is often able to predict the
discrepancy fairly well. In our simulations, we found the gap estimation to fail only in areas
which contain both highly dense and very sparse areas. In the femtocell application, the
urban case is the most relevant, as rural and suburban users would mainly connect to
macrocells and it appears that the PPP model is a good approximation for urban areas.
The SIR distribution inurban areas agreeswell with what the PPPmodel predicts.Suburban and ruralareas show somediscrepancy, but arewell approximatedusing gap estimation.
5
Figure 3: Comparison of the ccdf of SIR from the PPP model (red solid) and from themodel using real house locations (blue dots) (a) in an urban area (2km by 2km) and (b) in asuburban/rural area (10km by 10km). The black dashed and blue dot dashed curves showthe approximated SIR curve using gap estimation. Results from urban areas agree very wellwith the analytical formula for the PPP model. In suburban/rural cases, there is asignificant discrepancy which is approximated well by gap estimation. The further towardthe right the curve is, the better the performance.
4. Comparison with commercial softwareSo far, we have been using simplified signal propagation models, which may cause
inaccuracies. To validate the results with even more realistic data, we obtained data for the
SIR distribution of an urban region in London produced using Atoll, a complex
commercial software produced by the company Forsk. The software takes into account
shadowing due to elevation and buildings, and uses more sophisticated propagation models
in order to find the SIR distribution. We observe that the data from the Atoll simulations
agrees well with both the PPP model and the model using actual house locations (see
Figure 4), with the latter being closer to the Atoll data. There is some discrepancy between
the two methods on the low SIR end (left end), which is likely due to Atoll automatically
setting SIR values to -6.5 dB, if they fall below this value. This strong agreement suggests
that the PPP model is a valid approximation for finding the SIR distribution in urban areas.
a
b
The ccdf of SIR fromthe PPP model agreeswell with SIR datafrom an urban areafrom simulationsproduced bycommercial software.
6
Figure 4: Comparison of the PPP model (PPP, red solid) and model with real houselocations (house model, blue dots) with data from commercial software (Atoll, blackdashed).
5. Inter-cell interference coordinationIn Figure 4, we see that, without any additional mechanism to improve the performance of
the system, the SIR for more than 40% of UEs is below 0 dB, a regime in which
connectivity to the network is typically impossible. For this reason, it is critical for network
providers to improve the SIR and with that the performance of the system. One such way
to improve the SIR is to use inter-cell interference coordination (ICIC). Under ICIC, cells
close together use different frequency spectra or time slots, reducing the interference to
UEs that receive signal from either of them. In practice, this reduces the interference from
the strongest interferers for UEs that would otherwise achieve very low SIR.
Figure 5: Performance of 2-ICIC compared to no-ICIC, when P1 = 0.5. We see that theaverage gain (horizontal shift) for UEs is only about 2 dB. This indicates that 2-ICIC is notlikely to improve the performance much.
We derived an analyticalformula for oneimplementation of ICIC,in which the interferencefrom nearest interferersis reduced.This formula can be usedpredict performance ofICIC for different powerlevels of the nearestinterferers.
7
In the previous sections, we have shown that the PPP model with Rayleigh fading can be
used for predicting the SIR distribution of urban areas. We test the performance of ICIC
by deriving analytical solutions to the ICIC case in a PPP model with pathloss and Rayleigh
fading. We consider the case in which the nearest interfering cell transmits at reduced
power P1, which represent the best-case scenario for a 2-cell coordination (2-ICIC). In
Figure 5, we see that the average gain is about 2 dB (as the curve is shifted to the right by
2 dB) in 2-ICIC when the nearest cell transmits at half power and we also compare our
analytical solution with the case without ICIC. The formula can be extended to the case
when the next nearest interferers are also at reduced powers, and is useful to test out
different scenarios of ICIC.
6. Conclusion and recommendationsWe have shown that the SIR distribution from PPP models agrees very well with more
realistic simulations obtained from both commercial software and by using actual house
location data for urban areas. For suburban and rural areas, there is a significant
discrepancy between the SIR distribution from the PPP model and these other approaches.
However, the gap (or shift) from the PPP model can often be approximated using results
derived in [2]. Urban areas are the main application area for femtocells, so our results
indicate that PPP models can be used to predict the SIR distribution for femtocell
deployment. Furthermore, as an example of how PPP models can be applied, we derived
analytical formulas for ICIC in a PPP model, which can be used to predict the
performance of a wireless system with ICIC.
Therefore, we would recommend using PPP models to predict the performance of a
femtocell deployment in urban areas under different scenarios such as ICIC or other
transmission coordination schemes. In a PPP model framework, analytical formulas for the
SIR distribution can be found for many different scenarios, such as ICIC.
7. Potential ImpactWe have found that PPP models can be used as first-order approximations of a femtocell
system in urban areas and hence are useful for finding the potential performance of a
femtocell system without needing to rely on complex simulations. The analytical formula
for ICIC that we derived is useful to find the performance gain using a given ICIC set-up
and extends the existing literature on ICIC.
Dr. Keith Briggs, research mathematician at BT Technology, Strategy and Operations,commented: “The wireless research team at BT is very interested in ways of predicting the performance ofsmall-cell deployments, and the methods Fabian has developed and tested demonstrate that in many casessimple mathematical models will be sufficiently accurate for our purposes. Furthermore, the ICIC work haspotential to be considerably extended.”
References
1. J. G. Andrews et al (2011) A Tractable Approach to Coverage and Rate in CellularNetworks IEEE Transactions on Communications. 59, no. 11. pp. 3122—3134
2. R. K. Ganti and M. Haenggi (2015) SIR asymptotics in general cellular network models,IEEE Transactions on Communications 64 no. 3, pp. 1009—1013
PPP models are goodapproximations of realfemtocell systems. Wecan use PPP models totest out different systemset-ups such as ICIC.